广告
加载中

MiniMax发布M3.1 Flash大模型 单任务成本低至0.22美元

亿邦AI 2026-10-20 09:20
亿邦AI 2026/10/20 09:20

邦小白快读

EN
全文速览

总述:MiniMax于2026年10月19日发布M3.1 Flash推理大模型,性能提升、成本降低,普通用户也能以很低价格体验前沿AI能力。

1. 核心性能:智能指数42分,比前代M3提高13分;复杂推理能力提升最大,在人类终考评测中得分48%,超过DeepSeek和GLM同级模型;事实准确性测试正确率56%,明显优于竞品。

2. 速度与效率:运行速度是前代3倍,输出速度中位数稳定在每秒170token以上,最高超200token;支持100万token上下文,可以一次性处理长文本。

3. 价格和使用:API每百万输入token仅0.15美元,输出0.6美元,缓存输入0.015美元,单任务平均成本0.22美元,比主要竞品都低;可通过官网API、MiniMax Code、M Plan订阅使用,10月23日将开放开源权重。

4. 风险注意:该模型幻觉率达93%,在不确定时倾向生成错误内容而不是拒答,个人使用重要信息时一定要核对来源。

总述:M3.1 Flash在销售、营销、运营、客服等企业流程中表现不俗,且成本低、速度快,对品牌商来说是值得关注的AI工具选项。

1. 企业场景适配:Automation-Bench测试中得分62.2,略高于DeepSeek同级别,覆盖销售、营销、运营、客服、财务、人力等岗位,可用于自动化客服回复、营销文案辅助、运营数据分析等。

2. 成本优势与定价:单任务平均成本0.22美元,API输入每百万token0.15美元,输出0.6美元,缓存输入0.015美元,低于多个竞品,做高频内容生成或客服调用成本更可控。

3. 内容质量与风险:事实准确性测试正确率56%,高于GLM和DeepSeek,但幻觉率高达93%,代表模型在不知道答案时容易编造内容。品牌商若将其用于对外客服或营销内容,必须加信息校验和人工审核,避免错误信息伤害品牌形象。

4. 多模态与扩展:支持文本和图像输入,上下文窗口100万token,可处理长文档、产品手册,对未来品牌内容个性化有潜力。

总述:M3.1 Flash的发布给卖家带来低成本AI工具和效率提升机会,但需注意模型幻觉带来的运营风险。

1. 降本机会:官方API定价处于全球最低区间,单任务平均成本0.22美元,约为Gemini同级的六分之一;卖家若用AI处理客服、售后、商品描述生成等高频场景,可明显降低单次调用费用。

2. 提效亮点:输出速度中位数每秒170token以上,上下文支持100万token,能一次性分析长篇客户反馈或政策文件;在销售、运营、客服等企业流程测试中得分62.2,略超DeepSeek同级,适合自动化常见运营任务。

3. 合作与获取方式:API已在MiniMax官方平台开放,也可通过桌面端MiniMax Code和M Plan订阅接入;10月23日将开放开源权重,卖家有技术能力可自部署,进一步降低长期使用成本。

4. 风险提示:模型幻觉率高达93%,在不确定时会生成错误内容,卖家不能完全依赖它处理客户订单、库存或财务数据,建议配置人工复核或使用校验工具。

总述:对大模型有数字化需求的工厂,M3.1 Flash的代码、计算机操作和企业流程能力可作为自动化升级的参考方案。

1. 自动化操作启示:OSWorld 2.0操作系统操作基准得分61.7,显示模型能辅助操作计算机界面;网络安全漏洞挖掘基准CyberGym得分82.2,可用于检测生产系统的安全漏洞,提升数字化系统的可靠性。

2. 代码开发支持:代码基准DeepSWE v1.1得分73.7%,接近Claude Opus 5和DeepSeek一级水平,工厂可用它辅助MES、ERP等系统的二次开发、生成自动化脚本或排查代码问题。

3. 成本与部署:API输入每百万token0.15美元,单任务成本0.22美元,速度是前代3倍;支持100万token上下文,能读大量设备文档和操作手册;10月23日开源,方便工厂部署在内部环境,数据不出厂。

4. 风险注意:幻觉率93%,在机械参数、操作指令等关键场景必须加校验和人工确认,否则可能造成错误操作或安全隐患。

总述:M3.1 Flash在成本、速度和性能上的组合对服务商很有吸引力,同时其高幻觉率为增值服务创造了空间。

1. 行业趋势:模型价格持续下探,单任务成本0.22美元,输出token消耗仅1.89亿,低于同级别竞品,说明稀疏性架构等新技术正有效降低AI推理成本。

2. 客户痛点:企业客户使用AI最担心错误输出,M3.1 Flash幻觉率高达93%,几乎不会拒答,服务商在集成时应主动提供信息校验、检索增强、输出护栏和人工复核等解决方案,帮客户规避风险。

3. 技术解决方案:该模型支持文本和图像输入,上下文窗口100万token,适合构建企业知识库问答、长文档分析、自动化客服等应用;在代码和计算机操作测试中得分接近顶级模型,服务商可用于开发代码助手和RPA类产品。

4. 开源机会:10月23日将开放开源权重,服务商可以基于开源模型做私有化部署和行业定制,形成差异化服务。

总述:对AI平台运营方而言,M3.1 Flash的开放API和开源计划是引入生态的候选产品,但需管理好模型的合规和幻觉风险。

1. 平台接入价值:MiniMax开放API,提供每百万输入token0.15美元、输出0.6美元的定价,支持100万token上下文和图像输入,平台方可以低成本接入该模型,丰富模型广场供给。

2. 开源生态机会:10月23日将开放开源权重,平台可以提前布局模型托管、推理服务或二次开发模板,吸引开发者入驻;早期版本代号Space Bunny累计处理token超200万亿,证明模型经过大规模验证。

3. 运营管理重点:该模型幻觉率高达93%,平台需要设置内容安全过滤、输出审核、用户提示等机制,避免错误内容对平台产生负面影响;同时关注开源许可条款尚未公布,后续需跟进合规授权。

4. 差异化优势:运行速度是前代3倍,输出token消耗远低于GLM、Gemini和DeepSeek,意味着在平台上提供该模型可以降低单位算力成本,提升资源利用率。

总述:M3.1 Flash的发布展现了推理大模型在低成本化和性能提升上的新进展,也暴露出高幻觉率这一值得研究的产业问题。

1. 技术与评测:该模型智能指数42分,复杂推理在人类终考中超过DeepSeek,物理推理CritPt落后DeepSeek;事实准确性AA-Omniscience正确率56%领先竞品,说明不同评测维度各有优劣,需要建立更立体的模型评估体系。

2. 成本与架构:模型采用稀疏性架构,运行速度达前代3倍,输出token总量1.89亿,远低于竞品,单任务成本0.22美元,反映高效架构对推理经济性的影响,是成本结构研究的新案例。

3. 商业模式:API定价全球最低区间,加上即将开源权重(10月23日),MiniMax采取开放加低价策略,可能改变大模型商业竞争格局;Pro版本规划中,显示产品线分层打法。

4. 问题与建议:幻觉率93%且极少拒答,意味着当前模型在认知边界上存在缺陷。研究者可关注如何通过训练目标、对齐策略或外部校验机制降低幻觉,并为AI内容安全监管提供参考。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Overview: On October 19, 2026, MiniMax released the M3.1 Flash reasoning model, delivering better performance and lower costs—making frontier AI capabilities accessible to ordinary users at a very affordable price.

1. Core performance: It scores 42 on the intelligence index, 13 points higher than the previous-generation M3. Its biggest gain is in complex reasoning, where it achieves 48% on a human-style final exam—exceeding DeepSeek and GLM models of the same tier. It also posts a 56% accuracy rate on factual-accuracy tests, clearly outperforming competitors.

2. Speed and efficiency: It runs 3x faster than its predecessor, with median output speed consistently above 170 tokens per second and peak speeds exceeding 200 tokens/sec. It supports a 1 million-token context, allowing long documents to be processed in a single pass.

3. Pricing and access: The API costs only $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached-input tokens. The average per-task cost is $0.22—lower than major competitors. It is available via the official API, MiniMax Code, and M Plan subscriptions, with open-source weights set for release on October 23.

4. Risk to note: The model has a 93% hallucination rate and tends to fabricate answers when uncertain rather than refusing to respond. For personal use of important information, always verify the source.

Overview: The M3.1 Flash shows strong performance across business flows such as sales, marketing, operations, and customer service, with low cost and high speed—making it a compelling AI option for brands.

1. Enterprise fit: It scores 62.2 on the Automation-Bench test, slightly above DeepSeek's equivalent model, covering sales, marketing, operations, customer service, finance, and HR. Use cases include automated customer-service replies, marketing-copy assistance, and operational-data analysis.

2. Cost advantage and pricing: The average cost per task is $0.22, with API pricing of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached-input tokens—lower than several competitors. This makes high-frequency content generation or customer-service invocation more budget-controllable.

3. Content quality and risk: It achieves 56% accuracy in factual-accuracy tests, higher than GLM and DeepSeek, but its hallucination rate is 93%, meaning it is prone to fabricating content when it does not know the answer. Brands using it for external customer service or marketing content must add information verification and human review to avoid brand damage.

4. Multimodal and scalability: It supports both text and image inputs and a 1-million-token context window, enabling processing of long documents and product manuals—an asset for future brand-content personalization.

Overview: The release of the M3.1 Flash presents sellers with a low-cost AI tool and efficiency gains, but also carries operational risks related to hallucination.

1. Cost savings: Official API pricing sits in the lowest global band, with an average cost per task of $0.22—roughly one-sixth of the comparable Gemini tier. Sellers using AI for high-frequency scenarios such as customer service, after-sales support, and product-description generation can notably reduce per-call costs.

2. Efficiency highlights: Median output speed is above 170 tokens per second, and the context length of 1 million tokens allows for analyzing long customer feedback or policy documents in one go. The model scores 62.2 on enterprise workflow tests covering sales, operations, and customer service—slightly above DeepSeek's equivalent—making it suitable for automating common operational tasks.

3. Access and options: The API is already available on MiniMax's official platform, and it can also be accessed through the desktop MiniMax Code and M Plan subscription. Open-source weights will be released on October 23, allowing sellers with technical capabilities to self-host and further reduce long-term costs.

4. Risk warning: With a 93% hallucination rate, the model tends to generate incorrect content when uncertain. Sellers should not fully rely on it for customer orders, inventory, or financial data, and should set up human review or use validation tools.

Overview: For factories with digitalization needs, M3.1 Flash's capabilities in coding, computer operation, and enterprise processes offer a reference for automation upgrades.

1. Automation-operations insight: It scores 61.7 on the OSWorld 2.0 operating-system benchmark, showing it can assist with computer-interface operations. On the CyberGym benchmark for cybersecurity vulnerability discovery, it scores 82.2, which can be used to detect security flaws in production systems and improve the reliability of digital infrastructure.

2. Code-development support: It scores 73.7% on the DeepSWE v1.1 code benchmark, close to the level of Claude Opus 5 and DeepSeek-tier models. Factories can use it to assist with secondary development of MES/ERP systems, generate automation scripts, or troubleshoot code issues.

3. Cost and deployment: API pricing is $0.15 per million input tokens, with an average per-task cost of $0.22 and 3x faster speed than the previous generation. It supports a 1-million-token context, making it possible to process large volumes of equipment documentation and operation manuals. It becomes open source on October 23, enabling factories to deploy it in internal environments without data leaving the facility.

4. Risk attention: The hallucination rate of 93% requires added verification and human confirmation for critical scenarios such as mechanical parameters and operation instructions; otherwise there is a risk of incorrect operation or safety hazards.

Overview: M3.1 Flash's combination of cost, speed, and performance is highly attractive to service providers, while its high hallucination rate also creates room for value-added services.

1. Industry trend: Model prices continue to decline—average task cost is $0.22, and output-token consumption is only 189 million, lower than comparable competitors. This shows that techniques such as sparse architecture are successfully reducing AI inference costs.

2. Customer pain point: Enterprise clients are most concerned about wrong outputs. M3.1 Flash has a hallucination rate of 93% and almost never refuses to answer, so service providers should proactively offer solutions such as information verification, retrieval-augmented generation, output guardrails, and human review to help clients mitigate risk.

3. Technical solutions: The model accepts text and image inputs and supports a 1-million-token context, making it suitable for building enterprise knowledge-base Q&A, long-document analysis, and automated customer-service applications. Its scores on code and computer-operation benchmarks are close to top-tier models, so service providers can use it to develop code assistants and RPA-style products.

4. Open-source opportunity: The open-source weights will be released on October 23, allowing service providers to make private deployments and industry-specific customizations to form differentiated offerings.

Overview: For AI platform operators, M3.1 Flash's open API and open-source plans make it a candidate for ecosystem integration, though compliance and hallucination risks need to be managed.

1. Platform value: MiniMax offers an open API with pricing of $0.15 per million input tokens and $0.60 per million output tokens, plus a 1-million-token context and image-input support. Platforms can integrate the model at low cost to enrich their model marketplace.

2. Open-source ecosystem opportunity: The open-source weights will be released on October 23. Platforms can prepare model hosting, inference services, or secondary development templates in advance to attract developers. The early version, codenamed 'Space Bunny', has processed over 200 trillion tokens cumulatively, proving the model's large-scale validation.

3. Operational management focus: The model's hallucination rate is as high as 93%, so platforms need to implement content-safety filtering, output review, and user prompts to prevent misinformation from having adverse effects. Note also that open-source license terms have not yet been announced and must be monitored for compliant licensing.

4. Differentiation advantage: It runs 3x faster than the previous generation and consumes far fewer output tokens than GLM, Gemini, and DeepSeek, meaning that offering this model on a platform can lower unit computing costs and improve resource utilization.

Overview: The release of M3.1 Flash demonstrates new progress in low-cost, high-performance reasoning models, while also exposing the high hallucination rate that deserves further study as an industry-wide issue.

1. Technology and evaluation: The model scores 42 on the intelligence index and outperforms DeepSeek in complex reasoning in a human-style final exam, though it lags DeepSeek on the physical-reasoning CritPt benchmark. It leads competitors with 56% accuracy on the AA-Omniscience factual-accuracy test, showing that different evaluation dimensions have varied strengths and weaknesses—reinforcing the need for a more comprehensive model evaluation framework.

2. Cost and architecture: The model uses a sparse architecture, runs 3x faster than its predecessor, consumes only 189 million output tokens—far less than competitors—and costs an average of $0.22 per task. This reflects how efficient architectures affect inference economics and offers a new case study for cost-structure research.

3. Business model: With API pricing in the lowest global band and the move toward open-sourcing weights on October 23, MiniMax is combining openness with low prices in a strategy that may reshape the competitive landscape for large models. A planned Pro version also indicates a layered product-line approach.

4. Issues and recommendations: A hallucination rate of 93%, combined with the model's near-total reluctance to refuse answers, highlights current limitations in model cognitive boundaries. Researchers can explore how training objectives, alignment strategies, or external verification mechanisms can reduce hallucination, and use these findings as a reference for AI content-safety regulation.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

2026年10月19日,中国人工智能企业MiniMax正式推出全新推理大模型M3.1 Flash。

独立第三方机构Artificial Analysis的测评结果显示,该模型在其智能指数测评中拿到42分,较前代产品M3提升13分。这一成绩与智谱GLM-5.3-Flash持平,略高于谷歌Gemini 3.8 Flash的41分,距离DeepSeek近期发布的V4.1 Flash 1017仅差3分,单任务测评成本较前代降低超一半。

复杂推理能力是该模型升级幅度最大的板块。在“人类终考”测评中,M3.1 Flash得分48%,超过DeepSeek V4.1 Flash 1017的43%与GLM-5.3-Flash的40%。覆盖71道研究级物理推理题的CritPt基准测试中,该模型得分22%,领先GLM-5.3-Flash的15%、Gemini 3.8 Flash的18%,仅落后DeepSeek V4.1 Flash 1017的24%。事实准确性维度,AA-Omniscience测试中该模型答题正确率为56%,明显高于GLM-5.3-Flash的28%与DeepSeek V4.1 Flash 1017的43%。

M3.1 Flash运行速度达到前代M3的3倍,这一效率提升部分来自模型采用的稀疏性架构设计,可减少生成回复所需的计算量。测评过程中该模型消耗的输出token总量约为1.89亿,远低于GLM-5.3-Flash的3.24亿、Gemini 3.8 Flash的3.34亿与DeepSeek V4.1 Flash 1017的2.24亿。更低的token消耗量拉低了实际使用成本,单任务平均成本约为0.22美元,低于GLM-5.3-Flash的0.28美元、DeepSeek V4.1 Flash 1017的0.26美元,对比Gemini 3.8 Flash 1.24美元的单任务成本优势更加明显。

面向企业工作流、代码开发、计算机操作等场景,M3.1 Flash的测评成绩同样具备竞争力。覆盖销售、营销、运营、客服、财务、人力等岗位企业应用流程的Automation-Bench v1.0.6测试中,该模型拿到62.2分,略高于DeepSeek V4.1 Flash 1017的60.5分。聚焦多步骤金融工作流的BankerToolBench测试得分76.2,与Claude Opus 5持平。代码基准DeepSWE v1.1测试中得分73.7%,接近Claude Opus 5的74%、DeepSeek V4.1 Flash 1017的77%。操作系统操作基准OSWorld 2.0得分61.7,网络安全漏洞挖掘基准CyberGym得分82.2。

测评同时暴露了该模型的明显短板。M3.1 Flash在测评中会尝试回答所有问题,遇到不确定的内容极少拒答,最终测得幻觉率为93%。该数据统计的是错误回答在所有非正确反馈中的占比,并非整体答案错误率,即模型无法给出正确答案时,绝大多数情况会生成错误内容,而非坦承自身知识边界。企业将其用于知识密集型工作流时,需要额外配置信息校验、检索系统、输出护栏或人工复核环节,规避错误输出带来的风险。

M3.1 Flash支持文本、图像两种输入形式,上下文窗口长度可达100万token,输出速度中位数稳定在170token每秒以上,最高生成速度超200token每秒。官方公布的API定价为每百万输入token 0.15美元,每百万输出token 0.6美元,缓存输入价格低至每百万token 0.015美元,与DeepSeek V4.1 Flash 1017非高峰时段定价持平,处于当前全球前沿大模型最低价格区间。目前该模型API已在MiniMax官方平台正式开放,用户也可通过桌面端应用MiniMax Code、M Plan订阅服务接入使用。

MiniMax计划在10月23日开放该模型的开源权重,暂未公布对应的开源许可条款。官方同时披露,该模型的早期版本曾以Space Bunny为代号匿名运行,累计处理token总量超过200万亿,更高性能的Pro版本已在规划中,具体发布时间尚未公布。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

MiniMax M3.1 Flash是什么?

M3.1 Flash是MiniMax于2026年10月19日发布的推理大模型,支持文本与图像输入,上下文窗口达100万token,输出速度中位数超170token/s。它在企业工作流、代码开发、计算机操作等场景中表现较强,API已开放,并计划于10月23日开源权重。

M3.1 Flash的API价格是多少?比竞品便宜多少?

官方API定价为每百万输入token 0.15美元、每百万输出token 0.6美元,缓存输入低至0.015美元。在测评中,M3.1 Flash单任务平均成本约0.22美元,低于GLM-5.3-Flash的0.28美元、DeepSeek V4.1 Flash 1017的0.26美元,明显低于Gemini 3.8 Flash的1.24美元。

M3.1 Flash有什么缺点?

M3.1 Flash的主要缺点是幻觉率偏高。在测评中,模型面对不确定内容时极少拒答,幻觉率达93%。在无法给出正确答案时,它多数会生成错误内容。企业用于知识密集型工作流时,需要额外配置信息校验、检索系统、输出护栏或人工复核,以规避错误输出风险。

M3.1 Flash在哪些测评中领先?

在Artificial Analysis智能指数测评中,M3.1 Flash得42分,超过Gemini 3.8 Flash的41分;在“人类终考”测评中得分48%,超过DeepSeek V4.1 Flash 1017的43%和GLM-5.3-Flash的40%;在AA-Omniscience测试中正确率56%,明显高于GLM-5.3-Flash的28%。在自动化流程和金融工具等场景基准中也与顶级模型持平。

M3.1 Flash的开源计划是怎样的?

MiniMax计划于2026年10月23日开放M3.1 Flash的开源权重,但尚未公布开源许可条款。该模型的早期版本曾以Space Bunny为代号匿名运行,累计处理token总量超过200万亿。更高性能的Pro版本已在规划中,发布时间未公布。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0